Artificial intelligence (AI) is the driving force behind innovation in countless sectors. But how do you protect these breakthroughs? This issue can be particularly complex in the United States.
For years, the exception for “abstract ideas” posed a high barrier; abstract ideas were not eligible for patent protection under the law. Since the infamous Alice ruling by the Supreme Court in 2014, many AI inventions have fallen under Section 101 of the US Patent Act based on the test laid down by the Supreme Court in its ruling. Based on this test, the reasoning of the United States Patent & Trademark Office (USPTO) was strict when assessing patent applications, and the USPTO’s judgment was often as simple as it was deadly: it is ‘just’ mathematics on a computer.
That now seems to be changing. With the recent, precedent-setting decision in Ex parte Desjardins (fall 2025), the USPTO is marking a crucial change of course.
Catastrophic forgetting
The crux of the matter revolved around a patent application by Google/DeepMind. They developed a solution for ‘catastrophic forgetting’. This is a well-known problem whereby an AI model loses knowledge of an initial task as soon as it is trained for a second task. DeepMind’s invention involves a specific training technique that protects essential parameters so that the model ‘learns’ better without ‘forgetting’.
To understand the impact of this ruling, we need to look at what happened behind the scenes. Interestingly, the original Examiner had not rejected the application on the basis of § 101 (patentability). It was the USPTO’s appeals body (the PTAB, Patent Trials and Appeal Board) that added this ground for rejection on its own initiative. The PTAB’s ruling was harsh. Despite the technical context, the invention was essentially seen as an abstract mathematical algorithm on generic hardware.
This is where the top brass at the USPTO stepped in. The case was reviewed by a special Appeals Review Panel (ARP). The panel overturned the PTAB’s § 101 rejection, and on 4 November 2025, this decision was declared binding precedent. This means that all Examiners and judges within the USPTO are required to follow this new, more flexible line.
Practical application
The panel acknowledged that the claims contained mathematical concepts. However, it ruled that these concepts were integrated into a ‘practical application’. Because the invention achieved a concrete improvement in the functioning of the AI model itself, the analysis stopped there. It was no longer necessary to search for an additional ‘inventive concept’, the subjective hurdle that had previously been the death knell for many software patents.
The core of the Desjardins decision is the emphatic reaffirmation of the Enfish doctrine. This doctrine states that software is patentable if it makes a specific improvement to the computer functionality itself. According to the description, solving ‘catastrophic forgetting’ led to reduced storage requirements and less system complexity. These are tangible, technical advantages, not abstract results.
The most important implication of Desjardins is the instruction to examiners to shift their focus. The focus must be on the technical content. This means that the battle is shifting. The discussion is no longer about whether software is patentable at all, but whether the specific invention is innovative enough compared to existing technology (cf. novelty and inventiveness).
Although the winds are blowing more favorably at the USPTO, precision remains essential. The success in Desjardins relies heavily on how the invention was described. A patent application must explicitly articulate how the AI improves computer performance, for example less memory, faster processing, more efficient architectures, etc. Claims must include specific algorithmic steps. It is therefore advisable to claim the specific mechanism by which the objective is achieved, rather than just the broad principle.
Desjardins binds the USPTO, but federal courts (such as the Federal Circuit) sometimes take an even stricter line, especially when AI is used for purely business applications (as in Recentive Analytics v. Fox Corp.). However, inventions that improve the AI infrastructure itself are stronger than ever.
Not mere mathematics
Desjardins is therefore a very positive development for tech companies. The USPTO now recognizes that innovations in AI algorithms are not “mere mathematics” but technical solutions that deserve protection.
We are happy to assist you in navigating this changing landscape and drafting AI patent applications that meet the latest standards. Please feel free to contact one of our specialized patent attorneys.
Read also our dossier ‘Patenting artificial intelligence (AI) inventions’
